merge with master

CHANGELOG.md

@@ -10,13 +10,17 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 
 ### Added
 - Added `--no-spm-encode` option, allowing the model to use vocabulary IDs directly to train/decode.
+- Added MSE and MAE costs to COMET-QE training.
+- Added augmentation of shuffled examples to COMET-QE training via `--comet-augment-bad`.
+- Minor changes and fixes related to metric training.
+- Added `--quiet-validation` option that disables printing Hyp/Ref samples during validation
 - Added --custom-fallbacks option that allows to specify a list of option sets that get traversed for subsequent fallbacks upon divergence
-- Added --overwrite-checkpoint option that (when set to false) can be used to dump checkpoints with iteration numbers.   
+- Added --overwrite-checkpoint option that (when set to false) can be used to dump checkpoints with iteration numbers.
 - Implementations of COMET-20 (reference-based) and BLEURT-20 for inference with conversion scripts.
 - `./marian evaluate` sub command for evaluation with COMET-QE-20, COMET-20 and BLEURT-20
 - A bunch of scripts for metrics use and early MBR experiments
 - LSH vocab filtering for GPU. Speed is not competitive with non-LSH. Checking in for completeness and possible future use of LSH on GPU for non-filtering stuff
-- Added --throw-on-divergence and --fp16-fallback-to-fp32 options to detect (fp16 and fp32) and recover (only fp16) 
+- Added --throw-on-divergence and --fp16-fallback-to-fp32 options to detect (fp16 and fp32) and recover (only fp16)
   diverged runs. If not recoverable, exception gets rethrown and goes unhandled to force fatal error and shutdown.
 - Re-implementation of COMET-QE for inference and training; conversion scripts from Unbabel-Comet to Marian.
 - Validator that generates embeddings and can be used during COMET training with an external script.

VERSION

@@ -1 +1 @@
-v1.12.12
+v1.12.14

azure-pipelines.yml

@@ -293,8 +293,9 @@ stages:
 
     # https://software.intel.com/content/www/us/en/develop/articles/installing-intel-free-libs-and-python-apt-repo.html
     - bash: |
-        wget -qO- "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB" | sudo apt-key add -
-        sudo sh -c "echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list"
+        sudo mkdir -p /usr/share/keyrings
+        wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/intel.gpg > /dev/null
+        echo "deb [signed-by=/usr/share/keyrings/intel.gpg] https://apt.repos.intel.com/mkl all main" | sudo tee /etc/apt/sources.list.d/intel-mkl.list
         sudo apt-get update -o Dir::Etc::sourcelist="/etc/apt/sources.list.d/intel-mkl.list"
         sudo apt-get install -y --no-install-recommends intel-mkl-64bit-2020.0-088
       displayName: Install MKL
@@ -414,8 +415,9 @@ stages:
 
     # https://software.intel.com/content/www/us/en/develop/articles/installing-intel-free-libs-and-python-apt-repo.html
     - bash: |
-        wget -qO- "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB" | sudo apt-key add -
-        sudo sh -c "echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list"
+        sudo mkdir -p /usr/share/keyrings
+        wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/intel.gpg > /dev/null
+        echo "deb [signed-by=/usr/share/keyrings/intel.gpg] https://apt.repos.intel.com/mkl all main" | sudo tee /etc/apt/sources.list.d/intel-mkl.list
         sudo apt-get update -o Dir::Etc::sourcelist="/etc/apt/sources.list.d/intel-mkl.list"
         sudo apt-get install -y --no-install-recommends intel-mkl-64bit-2020.0-088
       displayName: Install MKL
@@ -606,8 +608,9 @@ stages:
 
     # https://software.intel.com/content/www/us/en/develop/articles/installing-intel-free-libs-and-python-apt-repo.html
     - bash: |
-        wget -qO- "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB" | sudo apt-key add -
-        sudo sh -c "echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list"
+        sudo mkdir -p /usr/share/keyrings
+        wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/intel.gpg > /dev/null
+        echo "deb [signed-by=/usr/share/keyrings/intel.gpg] https://apt.repos.intel.com/mkl all main" | sudo tee /etc/apt/sources.list.d/intel-mkl.list
         sudo apt-get update -o Dir::Etc::sourcelist="/etc/apt/sources.list.d/intel-mkl.list"
         sudo apt-get install -y --no-install-recommends intel-mkl-64bit-2020.0-088
       displayName: Install MKL

scripts/comet/comet2marian.py

@@ -90,6 +90,8 @@ def load_comet_model(model_path):
     config["type"] = "comet"
 elif model_type == "ReferencelessRegression":
     config["type"] = "comet-qe"
+elif model_type == "XLMRobertaModel":
+    config["type"] = "comet-qe"
 else:
     raise Exception(f'Unknown type of model {model_type}')
 
@@ -104,8 +106,10 @@ def load_comet_model(model_path):
 config["bert-train-type-embeddings"] = False
 config["bert-type-vocab-size"] = 0
 config["comet-prepend-zero"] = True
-config["comet-final-sigmoid"] = args.add_sigmoid
-config["comet-pooler-ffn"] = [2048, 1024]
+if not args.roberta:
+    config["comet-final-sigmoid"] = args.add_sigmoid
+    config["comet-pooler-ffn"] = [2048, 1024]
+
 # @TODO: figure out if it's worth adding `cometModel.name_or_path` to the end of this version string.
 config["version"] = "comet2marian2.py conversion"
 config["enc-depth"] = 0

src/common/config_parser.cpp

@@ -334,11 +334,14 @@ void ConfigParser::addOptionsModel(cli::CLIWrapper& cli) {
 
   // Options specific for the "comet-qe" model type
   cli.add<bool>("--comet-final-sigmoid", "Add final sigmoid to COMET model");
+  cli.add<bool>("--comet-stop-grad", "Do not propagate gradients through COMET model");
+
   cli.add<bool>("--comet-mix", "Mix encoder layers to produce embedding");
   cli.add<bool>("--comet-mix-norm", "Normalize layers prior to mixing");
   cli.add<float>("--comet-dropout", "Dropout for pooler layers", 0.1f);
   cli.add<float>("--comet-mixup", "Alpha parameter for Beta distribution for mixup", 0.0f);
   cli.add<bool>("--comet-mixup-reg", "Use original and mixed-up samples in training");
+  cli.add<float>("--comet-augment-bad", "Fraction of bad examples added via shuffling for class/label 0.f", 0.0f);
   cli.add<std::vector<int>>("--comet-pooler-ffn", "Hidden sizes for comet pooler", {2048, 1024});
   cli.add<bool>("--comet-prepend-zero", "Add a start symbol to batch entries");
 
@@ -705,6 +708,11 @@ void ConfigParser::addOptionsValidation(cli::CLIWrapper& cli) {
       "Keep best model for each validation metric");
   cli.add<std::string>("--valid-log",
      "Log validation scores to file given by arg");
+
+  // general options for validation
+  cli.add<bool>("--quiet-validation",
+    "Suppress logging hyp./ref. samples during validation");
+
   cli.switchGroup(previous_group);
   // clang-format on
 }

src/data/corpus.cpp

@@ -143,7 +143,7 @@ SentenceTuple Corpus::next() {
       // check if all streams are valid, that is, non-empty and no longer than maximum allowed length
       if(std::all_of(tup.begin(), tup.end(), [=](const Words& words) {
             return words.size() > 0 && words.size() <= maxLength_;
-          })) {
+         })) {
         return tup;
       } else {
         return SentenceTupleImpl(); // return an empty tuple if above test does not pass

src/data/corpus_base.cpp

@@ -434,22 +434,34 @@ void CorpusBase::addWordsToSentenceTuple(const std::string& line,
     auto prependedWord = Word::fromWordIndex(0);
     words.insert(words.begin(), prependedWord);
   }
-
-  if(maxLengthCrop_ && words.size() > maxLength_) {
-    words.resize(maxLength_);
+
+  // if fields are joined and the current sentence is not the first one, we need to make sure that
+  // the current sentence is not longer than the maximum length minus the length of the previous sentence
+  // (minus 1 for the separator <eos> token)
+  size_t localMaxLength = maxLength_;
+  if(joinFields_ && !tup.empty())
+    localMaxLength = std::max(1, (int)maxLength_ - (int)tup.back().size());
+
+  // if the current sentence is longer than the maximum length, we need to crop it
+  if(maxLengthCrop_ && words.size() > localMaxLength) {
+    words.resize(localMaxLength);
     if(addEOS_[batchIndex])
       words.back() = vocabs_[batchIndex]->getEosId();
   }
 
+  // if true, the words are reversed
   if(rightLeft_)
     std::reverse(words.begin(), words.end() - 1);
 
   // if true, the numeric indices get joined with the previous sentence, <eos> acts as a separator here
-  // @TODO: make this cleaner.
-  if(joinFields_)
-    tup.appendToBack(words);
-  else
+  if(joinFields_) {
+    size_t currLength = tup.empty() ? 0 : tup.back().size();
+    // if the current sentence would exceed the maximum length we don't add any more fields
+    if(currLength + words.size() < maxLength_)
+      tup.appendToBack(words);
+  } else {
     tup.pushBack(words);
+  }
 }
 
 void CorpusBase::addAlignmentToSentenceTuple(const std::string& line,

src/data/corpus_base.h

@@ -117,6 +117,9 @@ class SentenceTupleImpl {
   auto begin() const -> decltype(tuple_.begin()) { return tuple_.begin(); }
   auto end() const -> decltype(tuple_.end()) { return tuple_.end(); }
 
+  auto begin() -> decltype(tuple_.begin()) { return tuple_.begin(); }
+  auto end() -> decltype(tuple_.end()) { return tuple_.end(); }
+
   auto rbegin() const -> decltype(tuple_.rbegin()) { return tuple_.rbegin(); }
   auto rend() const -> decltype(tuple_.rend()) { return tuple_.rend(); }
 

src/graph/expression_operators.cpp

@@ -362,27 +362,9 @@ Expr flatten_2d(Expr a) {
   return Expression<ReshapeNodeOp>(a, shape);
 }
 
+// Stop gradients from flowing back through this node
 Expr stopGradient(Expr a) {
-#if 0 
-  // This is a different implementation which is more reliable than the original, 
-  // but it introduces a full copy which hogs memory. Keeping it around for now
-  // to decide later which one to use.
-
-  auto fwd = [](Expr output, const std::vector<Expr> inputs) {
-    CopyCast(output->val(), inputs[0]->val());
-  };
-
-  auto bwd = [](Expr output, const std::vector<Expr> inputs) {
-    /*Dummy*/
-  };
-
-  return lambda({a}, a->shape(), a->value_type(), fwd, bwd, (size_t)&fwd);
-#else
-  // implemented as a dummy reshape that is not trainable
-  auto res = Expression<ReshapeNodeOp>(a, a->shape());
-  res->setTrainable(false);
-  return res;
-#endif
+  return Expression<StopGradientNodeOp>(a);
 }
 
 Expr choose(std::vector<Expr> nodes, size_t index) {

src/graph/node_operators_unary.h

@@ -1042,6 +1042,29 @@ class ClipGradientNodeOp : public UnaryNodeOp {
   }
 };
 
+// This class is used to stop gradients from flowing through a node.
+// This is slighyly costly, as memory is allocated and a copy is made on the forward step.
+class StopGradientNodeOp : public UnaryNodeOp {
+public:
+  StopGradientNodeOp(Expr a)
+   : UnaryNodeOp(a) {}
+
+  ~StopGradientNodeOp() {}
+
+  // On forward the values are just copied from the input node.
+  NodeOps forwardOps() override {
+    return { NodeOp(CopyCast(val_, child(0)->val())) };
+  }
+
+  // On backward nothing happens, i.e. the gradient does not flow through.
+  NodeOps backwardOps() override {
+    return { NodeOp( /*dummy*/; ) };
+  }
+
+  const std::string type() override { return "stopGradient"; }
+  const std::string color() override { return "grey"; }
+};
+
 // narrow an axis to [begin, end)
 // The resulting object must be consecutive in memory.
 class SliceViewNodeOp : public UnaryNodeOp {